Kalman filtering and subspace projection approach to multispectral and hyperspectral image classification

Linear unmixing is a widely used remote sensing image processing technique for sub-pixel classification and detection where an image pixel is generally modeled by a linear mixture of spectral signatures of materials present within the pixel. This approach operates on a pixel by pixel basis and assumes that the image data are stationary and pixel-independent. Unfortunately, this is not true for real data due to varying atmospheric and scattering effects. Kalman filtering is one of the most commonly used techniques in communications/signal processing to deal with nonstationary environments in real time processing. This dissertation presents an approach, called Linear Unmixing Kalman Filtering (LUKF) which incorporates the concept of linear unmixing into Kalman filtering to take advantage of interpixel correlation so as to achieve signature abundance estimation, subpixel detection and classification for remotely sensed images. In this case, the linear mixture model used in linear unmixing is implemented as the measurement equation in Kalman filtering. The process equation which is required for Kalman filtering, but is absent in linear unmixing, is then used to model the signal.;Since the developed LUKF assumes a complete knowledge of the signature matrix used in the linear mixture model, it is also extended to an unsupervised LUKF where no knowledge about the signature matrix is required a priori. The unsupervised learning method proposed is derived from a vector quantization-based clustering algorithm. It employs a nearest-neighbor rule to group potential signatures resident within an image scene into a class of distinct clusters whose centers represent different types of signatures. These clusters' centers are then used as if they were true signatures in the signature matrix ILUKF. In order to evaluate the effectiveness of ILUKF using these signatures, SPOT and HYDICE images were used for assessment. The results showed that the ILUKF can also effectively detect and classify targets at a pixel scale by nulling interference resulting from unknown signatures.